Drama 3: native Metal support for the pinned NVFP4 pruned checkpoint #1266

Problem

The Drama recipe now recommends plotarmordev/drama-3-NVFP4-Pruned@494db08481db952f834b2241bce63fb3dfa72554. Its pinned CUDA engine (plotarmordev/TensorFold@b1f4b9b4830c9bd43bf7c432cffe7dea21313537) supports the checkpoint, but has no Drama Metal backend. The existing upstream Drama Metal loader expects the full FP8-expert checkpoint, all 256 experts per layer, and BF16 dense/tower projections. A launcher-only Mac change would therefore be misleading.

Proposed scope

Add support to the existing native Drama Metal family, preserving the existing full-model path and keeping the recipe's CUDA pin unchanged:

• Read the checkpoint's actual tensor formats and compact expert metadata; validate dtype, shape, scales, kept expert IDs and routing before use. The legacy bsgemm-moe config string alone is not sufficient to identify these mixed-format tensors.
• Execute packed NVFP4 experts on Metal without converting/requantizing the checkpoint or expanding the routed experts to a full BF16 model.
• Handle the checkpoint's dense/tower projection formats explicitly, retaining documented BF16 activations and FP32 accumulation. Unsupported combinations must fail clearly, never silently substitute another checkpoint/backend.
• Support the reference-conditioned guidance needed by the recipe, with explicit platform capability and voice-format handling rather than accepting ignored CUDA-only flags.
• Keep the change opt-in through selecting this checkpoint. No installed service changes, speed claim, quality-equivalence claim, or hardware-wide qualification claim.

Verification plan

Before claiming support: deterministic malformed-format/compaction/refusal tests, synthetic Metal projection checks against an independent FP64 decoder/reference (including signs, scale boundaries and mixed/pruned layers), native builds, and isolated real-checkpoint speech checks. Preserve the full-model behavior and exercise repeated requests, reference/default transitions, streaming/cancellation and owned-process cleanup where affected. Publish exact source/checkpoint/tool pins, attempted checks, failures and untested platforms. CPU tests are not GPU qualification; a successful startup is not successful synthesis.

The recipe PR will point to the engine contribution and its retained verification. Implementation is being prepared; no Mac NVFP4 compatibility result is claimed yet.